Can one model handle all scene text editing tasks?

TextWand unifies three separate text editing jobs—removal, generation, and replacement—into one model. The key innovations are a positional encoding method (ORPE) that locks text to precise positions and matches style from examples, plus a suppression strategy (RAS) that cleanly erases text. Authors built TextWand-Bench, the first broad benchmark for this problem, and show their model beats both open and closed competitors on text accuracy, layout consistency, and image quality.